YOLO-ACE: A Vehicle and Pedestrian Detection Algorithm for Autonomous Driving Scenarios Based on Knowledge Distillation of YOLOv10
Yuchen Xie, Danfeng Du, Mengju Bi
Abstract
Vehicle and pedestrian detection are critical tasks in autonomous driving, and fast and accurate detection algorithms are of great significance for improving the safety and reliability of autonomous driving systems. This paper proposes an improved YOLOv10 algorithm, YOLO-ACE, based on knowledge distillation for vehicle and pedestrian detection in autonomous driving scenarios. First, a new Add-CGLU (Additive-Convolutional Gated Linear Unit) architecture is developed to replace the original C2f module in the backbone part. Then, a new FPSC (Feature Pyramid Shared Conv) module is proposed to optimize the original SPPF module. After that, the neck part is redesigned to propose a new EMBS (Efficient Multi-Branch Scale) pyramid network. Finally, a new DD (Double Distillation) strategy is customized to perform knowledge distillation on the overall model. Experimental results on the public dataset BDD100K show that the computational parameters of YOLO-ACE are reduced by 21.6%, FLOPs are reduced by 20.0%, and the model size is reduced by 19.5%. At the same time, the F1 Score increased by 4.9%, the mAP increased by 4.5%, and the running speed reached 70.9 FPS. YOLO-ACE provides a more efficient vehicle and pedestrian detection solution in autonomous driving scenarios, promoting further development of autonomous driving systems.